Feature discussion: coordinating agent

[cledoux95]

Problem Statement

I want to the ability to have a coordinating agent that can take a series of implementation plans and spin up tasks within the application for me. This comes because I want to leverage parallel code's individual task worktree functionality, but I find myself manually coordinating all of those.

I have an initial branch that I'm working on, but I want to have a discussion on how you think the best way to approach this is before I get too deep into an implementation.

Possible Solution

Proof of concept branch:
branch

[johannesjo]

Thank you for the suggestion. Imho this is best done by configuring (or prompting) claude (or whomever) to do this properly. But if you have a good way to do this from our end or to provide a nice UI for this, I am all ears! :)

[cledoux95]

Now a power user can create a task that coordinates other agents

Coordinated tasks appear as children of the coordinating task. If you exit the coordinator, the children are unparented and continue as normal tasks.

[johannesjo]

That looks good (would probably put it after the two others though). How would you make it work? Via mcp? Or just instructions to to the agents?

[cledoux95]

I ran into some structural MCP server issues that I've just resolved, I'll be adding more screenshots and an explanation of how the coordinating agent works

[cledoux95]

Here's a review of the general architecture

[cledoux95]

A couple things that I added during debugging:

1. MCP server status indicator on the left side (we can remove)
2. MCP server debugging - planning on removing

The major design decision is:

1. Checking the "Coordinating agent" in the task setup alters the UI, but really doesn't do much else. It just preprends an instructional prompt to the users prompt to help the agent (regardless of which brand of agent) use the parallel-code MCP to coordinate the tasks.

Branch is up to date

[johannesjo]

Thank you very much for the detailed walkthrough! Before we think about implementation, I'd like to understand the use case better, since it never came up for me personally tbh:

1. When you imagine the coordinator agent, what actions does it actually need to take?

For example:
- Create new tasks (with a specific prompt/branch)
- Monitor task status (running, exited, exit code)
- Read agent output from other tasks
- Send follow-up prompts to running tasks
- Close/merge completed tasks

2. Is this more of a "dispatcher" (spins up tasks and walks away) or a "supervisor" (actively monitors and intervenes)?

3. Could you walk through a concrete example of what you'd want the coordinator to do step by step?

I am trying to figure out the simplest way to support this – there might be a lighter-weight approach than a full MCP server that covers the important cases.

[cledoux95]

Sounds good. I wrote the code in this branch as a functional POC - no pressure to reuse, we can scrap all of it. I just wanted to work through the use case with something I could play with directly.

1. Agent responsibilities
   ✅ Create new tasks
   ✅ Monitor the tasks
   ✅ Listen for responses from managed tasks
   ✅ Handle follow up prompts
   🟡 It could do this, but I'm very defensive of my repo, I don't give agents merge privileges, and I don't want it closing tasks without my review

2. Supervisor. IMO if theres going to be a reason to add a feature like this, it should be very capable.

3. Here's the use case that got me started:

   1. I took a list of 49 feature requests from my product beta testers and generated 49 implementation plans with claude.
   2. I was manually setting up agents one by one with the same prompt but just editing one thing, e.g. "read plan 1.9", "read plan 2.0", etc.
   3. Then after discussing with the agents, reviewing the plan and UX, I rand Red/Green TDD on each one to get the code written and tested.

After about 20 plans, I figured there was a better way, and that led me to the coordinating agent. The coordinating agent is there to take a set of instructions like "take these 49 plans and set up an agent for each one, with no more than 5 running at the same time. Ask each agent to read the claude.md, /docs folder, implementation plan, and then ask me questions to clarify the UX. After that, ensure the agent understands to perform Red/Green TDD and mark the implementation plan as done when all code is written, tested, and approved by the user."

The coordinating agent is currently doing that for me right now 😄

[johannesjo]

Thank you very much for the explanation. Very interesting and a cool smart solution you came with! Kudos! I'd love to find a good way to include this without adding too much complexity and overhead. To me this feels like a potentially very complex feature if you consider all the edge cases. I also have a feeling that the problem you had and similar cases might always require kind of a very special treatment, so it might be hard for us to really to find a good general solution.

But I think the main problem for me personally is that I don't really haven't had the use case yet. So chances are might have a completely different feeling about this then.

So I'd say: If you see a good, reliable solid minimal way of including this now, then let's do it. Otherwise I'd lean into testing this more for you personally. You can also always fork the project and go even stronger into that direction (maybe you're on to something here). Potentially a plugin system is also something we might want to evaluate in the future.

[cledoux95]

Sounds good. I'm basically at the fork already, I have it up and running and already managing parallel tasks. I think the code could probably be reduced a bit, but there's probably no way around an MCP server. Do you want me to open a WIP PR so you can see the code change diff?

[johannesjo]

Yes, I would like to test this for myself!

[ASRagab]

Would love to see this get in the app!